Research On Recommendation System Based On Deep Learning

In the era of big data,the problem of “information overload” has become a new trouble in life.As an effective information filtering tool,recommendation system plays an increasingly important role in the real world.With the rapid development of artificial intelligence,deep learning has been widely applied to various fields of machine learning such as natural language process and image process,and has achieved remarkable success.It is of great significant to employ deep learning to recommendation systems.With deep learning,users' preferences could be better learned from multimodal data,and recommendation systems could be more capable to find users' need.With the rapid development of Internet,user-item rating data have become increasingly sparse,which greatly enhanced the difficulty of recommendation.In order to alleviate the problem of sparseness,deep learning and multimodal data are considered to be employed in this paper.On the one hand,more abstract latent features of users and items could be extracted by deep leaning from multimodal data.On the other hand,the interaction between users and items could be better modeled by deep learning based methods cause of the characteristic of deep learning.In order to alleviate the sparsity problem in recommendation systems,main contributions of this paper are as follows.(1)For the purpose of alleviating the problem of sparsity and obtaining more suitable distribution of users' features in the condition of sparse data,a model called Constrained Probabilistic Matrix Factorization based on Neural Network(CPMF-NN)is proposed.Since different users have different rating history,they should also have different feature distributions.With this in mind,CPMF-NN takes the influence of users' interacted items into consideration for users' latent features.At the same time,CPMF-NN extracts latent features of items from the description texts of items by CNN.Finally fusing latent features of users and items with multi-layer perceptron for better capturing the nonlinear structural characteristics of interactions between users and items and improve the accurate of predicted ratings.(2)In order to further alleviate the problem of sparsity,a model called Recommendation Based on Multimodal Information of User-Item Interactions(RBMI)is proposed.For the sake of getting more suitable latent feature representation of users and items in the conditionof sparse data,the changes of users' preferences and multimodal data are considered.Firstly,CNN is used to extract items' latent features from the description texts of items.And then,for each user,RBMI takes the latent feature vectors and the corresponding ratings of the items that the user have rated previously as the input of a long short-term memory network for learning dynamic latent representation of the user.Finally,the sigmoid function is employed to calculate the interaction probability between users and items.